Using Syntax to Learn Semantic: An Experiment in Language Acquisition with a Mobile Robot

Abstract

Children learn natural languages by hearing utterances while interacting with their physical environment. We investigate one aspect of language acquisition by similarly situated, embodied artificial agents - using information about syntax to learn linguistically relevant semantic features. The agent is assumed to have no innate knowledge of syntax, and instead leverages the weak information about syntax available in word co-occurrences. Similarity of context (i.e.the surrounding words) is used to hierarchically cluster words, with clusters corresponding to sets of words that are similar syntactically and, often, semantically. The goal is to identify semantic features captured by the clusters. The leaves of the hierarchy are individual words, which are semantically very specific, and movement up the hierarchy leads to less specificity. The results of an experiment are discussed in which human subjects generated unrestricted natural language utterances to describe the activities of a Pioneer1 mobile robot. The combination of word clustering on this corpus and a common subsequence algorithm applied to the time series of sensor values recorded by the robot made it possible for the Pioneer1 to learn a variety of semantic features. May 17, 1999